Spiking Neural Network with Linear Computational Complexity for Waveform Analysis in Amperometry

Area of Science:

• Neuroscience
• Computer Science
• Electrical Engineering

Background:

• Spiking Neural Networks (SNNs) offer bio-inspired computational models.
• Edge computing requires efficient and low-power neural network architectures.
• Amperometric techniques, like those using Carbon NanoTube (CNT) sensors, detect minute currents.

Purpose of the Study:

• To design a simplified Spiking Neural Network (SNN) architecture for time waveform analysis on edge devices.
• To reduce the computational complexity of SNNs for practical implementation.
• To demonstrate the network's capability in analyzing low-current signals from amperometric sensors.

Main Methods:

• Developed an SNN architecture inspired by diencephalon signal preprocessing and thalamus spiking models.
• Implemented a simplified SNN by minimizing synaptic connections and weight dispersion.
• Designed a network mapping and learning algorithm with linear dependency on pattern size.
• Tested the accuracy stability across different network sizes.

Main Results:

• Achieved significant reduction in SNN algorithm complexity.
• Demonstrated the SNN's ability to process currents below 100 pA, suitable for amperometric sensing.
• Successfully applied the SNN to analyze vesicle fusion signals using CNT sensors.
• Validated the stability of the accuracy parameter for varying network sizes.

Conclusions:

• The proposed SNN architecture offers an efficient solution for time waveform analysis in edge computing.
• The network's low-power requirements make it suitable for amperometric sensing applications.
• Further analysis of implementation costs for semiconductor structures is discussed.